Belief Revision
Overview
Belief revision asks how an agent should respond to new information: accept it, adjust existing beliefs, or reject it while avoiding inconsistency. Formal approaches include AGM and Katsuno–Mendelzon postulates, plausibility orderings, and dynamic epistemic logic; newer systems also use explicit evidence records, dependency graphs, and uncertainty-aware update rules. In AI, these ideas guide language-model agents handling conflicting sources, persistent memory, user preferences, plans, and multi-turn dialogue, with evaluations on tasks such as conversational memory, navigation, and information seeking. A central tension is how to remain open to credible corrections without letting noisy evidence, social pressure, or one mistaken update destabilize useful beliefs and downstream actions.
Latest papers
29 papers1 paper in the last four weeks, against 2 the four weeks before. 0.0% of all new papers.